Global Fields of The World (FTW)

# Fields of the World — Global

> The first global, wall-to-wall agricultural field-boundary dataset at 10 m resolution: 3.17 billion
> field polygons across 241 countries and territories for 2024–2025, produced by applying the PRUE
> field-boundary segmentation model (a U-Net with an EfficientNet-B7 encoder, trained on the Fields of
> The World benchmark) to cloud-free Sentinel-2 mosaics. Published openly under CC-BY-4.0 by Taylor
> Geospatial and collaborators (Microsoft AI for Good, ASU, WashU in St. Louis, Oregon State, Clark)
> on Source Cooperative. Paper: Robinson et al. 2026, "The first global agricultural field boundary
> map at 10 m resolution" (arXiv:2605.11055).

This file describes the dataset for machines and assistants. The repository hosts several products
(vector field boundaries, raster confidence/quality layers, Sentinel-2 mosaics, Zarr stacks). A STAC
catalog is being built incrementally; it currently covers the **prediction confidence & quality
layers** and will expand to the other products.

Storage note: the public URL base is `https://data.source.coop/ftw/global-data/`, physically backed by
the S3 prefix `s3://us-west-2.opendata.source.coop/tge-labs/ftw-global-data/` (anonymous read).

## Key facts

- 3.17 billion field polygons (1.62 B in 2024, 1.55 B in 2025); 241 countries/territories; 10 m resolution.
- Model: PRUE (U-Net / EfficientNet-B7), trained on the CC-BY subset of Fields of The World (24 countries).
- A field here is a *remote-sensing field unit* (a connected component of predicted field-interior
  pixels), not a cadastral/legal parcel. The dataset is NOT a land-tenure product.
- Outputs are fiboa-compliant GeoParquet (vectors) and Cloud-Optimized GeoTIFFs (rasters).
- Validation: mean pixel-level recall 0.85 over 24 countries (14 > 0.90); confidence-model LOCO mean AUC 0.842.

## STAC catalog

- Root catalog: https://data.source.coop/ftw/global-data/catalog.json
- Collection — Prediction Confidence & Quality Layers (500 m):
  https://data.source.coop/ftw/global-data/predictions/confidence/collection.json
- Human README: https://data.source.coop/ftw/global-data/predictions/confidence/README.md

## Prediction confidence & quality layers (500 m)

Global 500 m rasters that quantify where the field predictions can be trusted. Common grid: EPSG:4326,
86400×34560 px, bbox [-180, -60, 180, 84], ~0.004167°/px; each 500 m cell ≈ 50×50 of the 10 m model
pixels (so pixel-count bands cap at 2500). Temporal coverage inferred as the 2025 prediction year
(pending author confirmation); the cropland-consensus layer is year-independent. Five STAC items:

- **confidence** — Modeled confidence score (Random Forest on model-internal indicators). Float32,
  nodata −1, range 0 to ~0.578. Recommended default filter conf ≥ 0.4; conservative conf ≥ 0.5.
  Item: .../predictions/confidence/confidence/confidence.json
  COG: .../confidence/prue_v1_confidence_global.tif (+ uint8 EPSG:3857 web-display variant).
- **field-density** — Field & boundary pixel counts per 500 m cell. uint16 (bands field_pixel_count,
  boundary_pixel_count). Variants: unfiltered, conf ≥ 0.4, conf ≥ 0.5, a default "filtered" product,
  a single-band fields-only, and a uint8 EPSG:3857 web-display variant.
  Item: .../predictions/confidence/field-density/field-density.json
- **entropy** — Mean Shannon entropy of the model softmax for the field and field-boundary classes
  (model uncertainty). Float32, 2 bands (mean_entropy_field, mean_entropy_boundary).
  Item: .../predictions/confidence/entropy/entropy.json
- **crop-consensus** — Mean agreement of 8 independent global cropland datasets (range 0–8; max 7
  outside Africa). External reference, year-independent. Float32, nodata 0.
  Item: .../predictions/confidence/crop-consensus/crop-consensus.json
- **precision-recall** — Per-cell precision/recall of fields vs the cropland-consensus agreement layer,
  at agreement ≥ 2 datasets (bands *_gt1) and ≥ 3 datasets (bands *_gt2). Float32, 4 bands.
  Item: .../predictions/confidence/precision-recall/precision-recall.json

Visualization styles (MapLibre GL, matching the FTW inference app palette via titiler):
.../predictions/confidence/styles/confidence.json (RdYlGn, rescale 0–0.578) and
.../predictions/confidence/styles/field-density.json (magenta→green).

## Data access

```python
# Read a window of the confidence COG (only the needed bytes are fetched)
import rasterio
from rasterio.windows import from_bounds
url = "https://data.source.coop/ftw/global-data/predictions/confidence/confidence/prue_v1_confidence_global.tif"
with rasterio.open(url) as ds:
    conf = ds.read(1, window=from_bounds(2.0, 47.5, 3.0, 48.5, ds.transform), masked=True)
```

```python
# Load the whole collection lazily as an xarray stack
import pystac, odc.stac
col = pystac.Collection.from_file(
    "https://data.source.coop/ftw/global-data/predictions/confidence/collection.json")
ds = odc.stac.load(list(col.get_items()), chunks={})
```

Browser preview (titiler): https://titiler.xyz/cog/preview.png?url=<COG-URL>&rescale=0,0.578178&colormap_name=rdylgn&nodata=-1

## Caveats

- Temporal year (2025) is inferred from the paper and pending author confirmation.
- The field-density "_filtered" variant's exact threshold/method is pending author confirmation.
- The confidence layer is conservative outside the FTW training distribution (e.g. smallholder
  systems): real fields there may get low confidence. Use the unfiltered density + continuous
  confidence rather than the default 0.4 threshold in such regions.
- Polygons are remote-sensing field units, not legal parcels; one parcel may map to many polygons or none.

## Other products in this repository (not yet in the STAC catalog)

- Vector field boundaries (fiboa GeoParquet), admin-partitioned by country:
  https://data.source.coop/ftw/global-data/predictions/vectors/alpha/results-by-admin/
- Raw polygonized predictions and PMTiles: .../predictions/vectors/alpha/
- Prediction probability Zarr: .../predictions/zarr/alpha/global.zarr
- Sentinel-2 feature mosaics (COGs + Zarr): .../features/cogs/alpha/ and .../features/zarr/alpha/global.zarr

## Cite

Robinson, C., Muhawenayo, G., Khanal, S., Fang, Z., Corley, I., Tárano, A. M., Estes, L., Marcus, J.,
Jacobs, N., Kerner, H., Becker-Reshef, I., & Lavista Ferres, J. M. (2026). The first global
agricultural field boundary map at 10 m resolution. arXiv:2605.11055.